Closed Sun-god-pineapple closed 2 years ago
we compute dx/dy by RANSAC plane fitting offline and implemented on cuda.
which described in paper 4.2. Propagation Loss :
We use the ground truth disparity d^{gt} and ground truth disparity gradients d{x}^{gt} and d{y}^{gt}, which we compute by robustly fitting a plane to d^{gt} in a 9x9 window centered at the pixel.
we compute dx/dy by RANSAC plane fitting offline and implemented on cuda. which described in paper 4.2. Propagation Loss :
We use the ground truth disparity d^{gt} and ground truth disparity gradients d{x}^{gt} and d{y}^{gt}, which we compute by robustly fitting a plane to d^{gt} in a 9x9 window centered at the pixel.
Thank you so much for your reply! I still have some questions down here.
- What is the meaning of parameter "sigma"?
sigma is a hyperparameter in RANSAC. that samples which errors less than sigma as "inliers".
- I found that you choose the defferent "iter" and "sigma" for SceneFlow Dataset and KITTI Dataset, could you please explain how you chose them ?
- I am trying to use your codes on another dataset, but some ground truth disparities in that dataset are very sparse. For example, some disparity may only have 10% of the pixels that have ground truth, while the remaing pixels have a value of 0. In that case, how can I set the parameters of the plane fitting so that the model can train good results. Looking forward to your reply at your convenience!
we tune these hyperparameters by observe quality of the visualized dx/dy. you can try this step by your self and adjust hyperparameters. i think too sparse ground truth can not get a good dx/dy estimation.
Thank you so much for your kind reply, It is really helpful! Wishing you a beautiful day!
Hi, Thank you so much for your wonderful contribution! I'm having some trouble understanding "plane_fitting_sf.py", especially this function down here Could you please explain this function for me and what part of the paper it corresponds to at your convenience? Thank you for your time!